LLMs are stuck in a groupthink groove. This startup is trying to get them out.

Editor's take

Large language models, when prompted for seemingly random numbers, exhibit a predictable bias, consistently returning specific digits like 7, 3, 4, and 8. This phenomenon suggests a shared underlying pattern in their training data or internal generation mechanisms, rather than true randomness.

The implication is that current LLMs may be less creative or independent than commonly assumed, potentially leading to a "groupthink" effect in applications requiring novel outputs. This is particularly relevant for fields like creative writing, drug discovery, or even generating diverse test cases, where predictable outputs limit their utility.

Future research should focus on understanding the specific data correlations or algorithmic choices that lead to this numerical predictability. Observing whether newer architectures or different training methodologies can produce genuinely unpredictable sequences, and how this impacts downstream tasks, will be key to assessing the extent of this LLM limitation.